Artificial intelligence still has some way to go

you guys know what MNIST is? it's 9.5MB of data. it would cost ~$100,000 to put that on the blockchain. so quite aside from the fact that the cryptocurrency applications are a scam, the actual practical distributed storage situation is ... not good.

Someone compiled a list of instances of AI doing what creators specify, not what they mean: https://t.co/OqoYN8MvMN pic.twitter.com/I1bgCFfO8c

— Jim Stormdancer (@mogwai_poet) November 7, 2018

this is why you should never tell your robot to cure world hunger

http://static.tvtropes.org/pmwiki/pub/images/tumblr_lm77uuW5G71qb98uxo1_500-2_dragged_6225.jpg

on the specific issue of deepfakes starting a war, i didn't find this essay as concretely comforting as i hoped i would, but it was interesting...

https://reallifemag.com/faked-out/

As long as mass media has existed in the West, there have been complaints about social acceleration, uncertainty, and the loss of a real, knowable world. In other words, our current conversations about the loss of reality are familiar; while each writer attempts to sound innovative, the concerns are evergreen. If the term “infocalypse” is useful, it is as a synonym for modernity, where truth is always two decades ago and dying today, and a new dark age always on the horizon.

i guess the two things i think are

1) infocalypse looks better on a page, but infopocalypse rolls off the tongue a little easier

2) as the article itself says near the end:

This analysis echoes philosopher Georges Bataille’s notion of “nonknowledge,” that the creation of knowledge always implies a corresponding creation of new fields of ignorance. Every revolution in information is also a revolution in misinformation. Information isn’t a light that shines down to answer questions; it also produces new unknowns, and possibilities for the unknown. Giving people more data and more information is as likely to create new readings and narratives than to align or “correct” beliefs. As Rob Horning recently put it, “any information, no matter how damaging it may seem to a particular side, can be put to any political use by any side; in fact no fact has any intrinsic meaning.” With each of modernity’s new technologies we are brought to recognize anew that belief precedes data rather than follows from it.

i don't know how to quantify the amount of information that is out there. not only the amount, but our ability to create it and share it, at a level of quality where an invented lie can more or less pass as something that could be true. but it seems like at least a metric shit ton to me. concerns about truth (and the ability capability to manipulate it) aren't new, but the potential scale for abuse seems much larger (and worrisome) now than ever before.

^^^

Someone took the neural net Yahoo trained to recognize NSFW pics and ran it "backwards" to generate images:

https://open_nsfw.gitlab.io/

whoa

the desert ones are really good

i was just gonna say that. very surreal

these are like, no based on specific images right? just representations of the schema the AI generated to recognize deserts, beaches, genitalia, etc? and that is based on like identifying patterns across billions of images right?

so we are looking into the dreamwork of AI, i guess. if i recognize what is happening, which i may not because this involved equations.

i think that's generally, right, although i must admit i skipped past all the text and went straight for the desert penis pics

god that shit is horrifying, reminds me of the artwork for Chris Cunningham's Rubber Johnny

it's extremely uncanny. a machine's vision of pornography, which doesn't even render actual bodies.

this is from someone with a dog in the race, and the numbers are shaky, but the basic point seems sound to me: no autonomous vehicles any time soon, except in very simple situations (or, if the super rich are able to swing it, unless we establish legal no go zones for non-autonomous vehicles)

https://medium.com/may-mobility/the-moores-law-for-self-driving-vehicles-b78b8861e184

That's a nice combination of simplified explanation and corporate self-promotion. I bet it is based on a presentation that CEO has given to dozens of VCs and banks.

I don't doubt any of it, but it does seem like a pretty convenient way for him to say essentially "It's okay that my company's cars are bad at autonomous driving (and aren't getting better fast enough), because everyone else's are too!"

yes, stipulated in my post (and his!). doesn't change the fact that he's right: it will be many decades before self driving cars actually work in the way typically claimed in the human environment we live in today.

I realize that, if this article were attempting to be rigorous, it wouldn't derive a 'Moore's Law for Autonomous Vehicles' by citing two data points for autonomous cars and one data point for human-driven cars and graphing them. It's miles from any kind of rigor. But what it suggests is more convincing to me than the claims that such fully autonomous vehicles will be arriving much sooner.

Between human performance (10⁸ miles per fatality) and the best-reported self-driving car performance (10⁴ miles per disengagement) is a gap of 10,000x. Put another way, self-driving cars are 0.01% as good as humans.

AYFKM

This ... only works if every disengagement would have resulted in a fatality.

I mean this is all messy stuff, I'm not saying his time estimates on general self-driving are wrong. But if Waymo removed its safety drivers early I think we'd see a lot more stuck vehicles than fatalities.

trump otm

In conversations on Air Force One and in the White House, Trump has acted out scenes of self-driving cars veering out of control and crashing into walls. https://t.co/RMUC3kcXKR

— Jonathan Swan (@jonathanvswan) March 17, 2019

lukas mentioned it above, but that medium article makes an incredibly disingenuous, apples-to-abstract-photography argument.

Between human performance (10⁸ miles per fatality) and the best-reported self-driving car performance (10⁴ miles per disengagement) is a gap of 10,000x. Put another way, self-driving cars are 0.01% as good as humans.

this is the graph that accompanies to quote:

https://i.imgur.com/POKRZWK.png

if we're asking self-driving cars to encounter a "disengagement" ("when the technology fails and a safety driver must take over") as often as a human being in a car has a fatality, then yes, it's going to be a while. but i'd argue that a better comparison would be human accidents to self-driving accidents. in the U.S., drivers have accidents roughly every 165,000 (a little over 10 to the 5th power). it's still an unfair comparison to compare human accidents to self-driving "disengagements", but even doing that makes things look quite a bit different:

https://imgur.com/IJ5wuxA

and just for fun, maybe it would be useful to compare self-driving "disengagements" to "human fuck ups". how often have you been in a car when the driver, perhaps yourself, did something catastrophically dumb but you escaped physical harm? personally, i have done so hundreds of times. i have had 1 accident in my life, and it was a low speed one and it was 99% the fault of a very confused elderly woman in a parking lot. anyway, i think i can conservatively estimate that at minimum, human drivers "fuck up" but escape physical harm 10 times more often than they have accidents. so that's at least an order of magnitude less frequent than accidents, or 10 to the 4th:

https://i.imgur.com/07LZsJo.jpg

so if the premise of the medium author's argument is right, self-driving cars should start having "disengagements" less frequently than human fuck-ups sometime a few years ago. i have no idea if that's true, but i think more than anything it suggests the whole argument is off

Another point to consider is that if a significant fraction of cars were driving autonomously, the driving AIs might well communicate to each other, or with Google maps or w/e, so that driving speeds and routes might be chosen so as to avoid risky situations altogether.

I hate AI. Kill it.

the driving AIs might well communicate to each other, or with Google maps or w/e, so that driving speeds and routes might be chosen so as to avoid risky situations altogether

Can you imagine how much chaos could be created by hacking into that network?

Doesn’t mean they won’t make it

Or fail to protect it adequately.

I don’t know what to tell you if you think self driving cars coordinating is a good idea in the present security environment.

Indeed, but done right it could seed traffic with a kind of "swarm intelligence" so that autonomous vehicles were acting in ways which improved overall flow. It sounds a little creepy, but I have no problems with traffic lights being managed to do the same thing, or routers managing the network traffic I use.
The collabortion (decided to let the typo stand) between Boeing and the FAA doesn't fill me with confidence though.

remember the robot bees yall

remember _KILLER_ bees yall, like when they were gonna be something that killed everyone

not everyone. only the poor sods who lived far enough south to allow Africanized bees to survive the winter.

so are they still a problem? i dont recall hearing a word about killer bees for at least a decade.

If the NY Post no longer pays attention to them, then I guess they've proved to be a manageable problem.

the killer bees murdered all the other bees and then went into hiding

Background: Last night, Google Maps' pricy contract with Zenrin, the high resolution map service in Japan, ended.

Google's cost savings solution was to implement their AI based on user behavior to build their new maps (what could possibly go wrong?):

https://www.reddit.com/r/japan/comments/b44wwt/so_many_people_were_cutting_across_the_konbini/

mountain shadows become lakes:

Googleマップが劣化したらしいと聞いて、近所の地図見たら、山影が湖になってたw pic.twitter.com/B237RUOpPA

— りん (@rin_kawakoubou) March 22, 2019

roads travel through/into buildings:

線路上や路上に建物あったり… pic.twitter.com/c2dqP1065D

— りん (@rin_kawakoubou) March 22, 2019

somewhat tangentially related: https://onezero.medium.com/how-googles-bad-data-wiped-a-neighborhood-off-the-map-80c4c13f1c2b

Apple maps and NAVITIME have have been better in japan for some time now, this definitely explains a few things.

NLP's moving so fast that benchmarking tasks introduced only a year ago are having to be replaced because they're running out of headroom: https://med✧✧✧.c✧✧✧@w✧✧✧.a✧✧✧.✧/introducing-superglue-a-new-hope-against-muppetkind-2779fd9dcdd5

uhhh maybe https://med✧✧✧.c✧✧✧@w✧✧✧.a✧✧✧.✧/introducing-superglue-a-new-hope-against-muppetkind-2779fd9dcdd5

well, I've run out of shits to give about getting this link to post correctly

looks like we won't be running out of headroom here anytime soon

Oh, I've got tons of empty space up there

https://medium.com/wang.alex.c/introducing-superglue-a-new-hope-against-muppetkind-2779fd9dcdd5

thanks

eh if NLP was solved then you'd notice in real life

good article
https://www.newyorker.com/magazine/2019/04/15/the-age-of-robot-farmers

auto-google-translating japanese post-article clickbait

500,000 yen-36 million yen! Benefits that may be obtained from the country!
Bad breath can be taken away! ? The bad breath care for seeing the cause stain was great
"Toughness ..." a husband who won't break his wife too? 980 yen is great now!
"It's a lie ..." Blood neutral fat, about 20% lower? What is the strength of Tokuho